How organisational context grows.
Published 2026-05-26 · 7 min read
A useful agent memory system is not a larger notebook. It is a feedback loop: capture work, constrain access, notice disagreement, ask a human when judgment matters, then let that answer improve the next retrieval.
The useful unit is not a transcript. It is a decision the next agent can safely act on.
Context has a lifecycle
Most agent memory systems begin with recall: save something, then retrieve it later. That is necessary, but it is not enough for a team. Work changes. Policies get replaced. Project exceptions appear. A private note becomes a shared convention, or a shared convention turns out to be too broad.
Ambience treats organisational context as something that is continually shaped by use. Every memory can become evidence, every conflict can become a review surface, and every review can update the organisation's working model.
01
Capture
An agent session leaves behind decisions, conventions, failures, and project facts worth reusing.
02
Constrain
Each memory gets scope, provenance, redaction state, and access checks before it can shape future work.
03
Extract
Durable claims are derived from the memory so the system can compare what changed over time.
04
Compare
Related claims are checked for disagreement, supersession, scope mismatch, or policy mismatch.
05
Review
Humans choose the current answer, keep both by scope, or dismiss a false positive.
06
Learn
The organisation's preference model is updated so the next recommendation starts closer to how the team works.
The memory remains the source of truth
The canonical record is still a memory. It carries enough structure to be governed: name, description, tags, scope, source, and audit history. It can be personal, team-level, project-level, organisation-wide, or sensitive.
That matters because retrieval is not just a ranking problem. A future agent can only receive context that the current user and agent are allowed to see. If a memory is superseded, flagged, or outside the caller's access boundary, it does not become ordinary context just because it matches a query.
Claims make memories comparable
A memory is written for people. A claim is written so the system can compare ideas. From each durable memory, Ambience extracts a small set of claims: the subject, the assertion, relevant qualifiers, where it applies, and any explicit time boundary.
This claim layer is intentionally derived. If the memory changes, the old claims are marked stale and new claims are generated. That keeps review auditable while allowing the model to reason over smaller, clearer units than a whole note.
Memory
Source memory with scope, provenance, redaction state, and audit history.
Claims
Structured assertions with applicability, confidence, and time boundaries.
Conflicts
Reviewable disagreements with recommendation, rationale, and status.
Org model
Team preferences about authority, scope, and repeated review reasons.
Conflicts are the learning moment
When two pieces of context disagree, the system should not bury the weaker one and hope nobody notices. It should show the disagreement, explain the recommendation, and make the human choice easy to inspect later.
Temporal validity
Which statement is current?
A launch plan from March conflicts with the rollout decision from May.
Factual correctness
Which statement should agents believe?
Two memories disagree about the approved retention policy.
Contextual applicability
Where is each statement true?
A default process is valid for the org, but one project has a deliberate exception.
Review trains the organisation model
A conflict resolution is more valuable than a thumbs-up. It says what the organisation trusts in practice: recent decisions over old plans, project exceptions over broad defaults, policy memories over one-off customer asks, or sometimes both memories because they apply in different places.
Accept suggestion
feedbackThe proposed winner becomes the preferred context and older memories can be superseded.
Choose other
feedbackA reviewer corrects the model's ranking and teaches the system what it missed.
Keep both
feedbackThe conflict is real, but both claims are useful when separated by time, scope, or customer.
Dismiss
feedbackThe system records a false positive instead of quietly pretending it was right.
Over time, those choices build an organisation profile: which scopes tend to win, which memory types are treated as high authority, which reason codes repeat, and where the system should be more careful next time.
What agents see
Agents do not need every internal detail of the conflict system. They need the right behavior at retrieval time.
Safe context that should carry its scope note with it.
Useful context, but the agent should see that the memory is contested.
Reviewable by humans, hidden from ordinary search results until resolved.
Some conflicts should appear directly in retrieval as a warning. Some are useful everywhere because scope is the key information. Some should stay in the Vault until a reviewer resolves them. The visibility choice is part of the conflict record, not an afterthought in the prompt.
Why this compounds
The result is a memory system that gets sharper through ordinary work. Nobody has to maintain a perfect handbook. The organisation improves its shared context by resolving the specific moments where an agent would otherwise be unsure.
That is the difference between a passive archive and an active context plane. An archive stores what happened. A context plane helps the next agent decide what to do with it.